[XPU] Fix/kimi linear xpu (#34546)

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Singh <rohitsi2@iil-login.iind.intel.com>
Co-authored-by: Singh <rohitsi2@iil-gnrap02.iind.intel.com>

python/sglang/srt/layers/attention/linear/kernels/kda_triton.py

@@ -5,7 +5,7 @@ import torch
 from sglang.srt.layers.attention.linear.kernels.kernel_backend import (
     LinearAttnKernelBase,
 )
-from sglang.srt.utils import is_cpu, is_npu
+from sglang.srt.utils import is_cpu, is_npu, is_xpu
 
 if not is_cpu():
     from sglang.kernels.ops.attention.fla.fused_recurrent import (
@@ -23,7 +23,10 @@ if not is_cpu():
 class TritonKDAKernel(LinearAttnKernelBase):
     """Triton-based kernel for KDA (Kimi Delta Attention) linear attention."""
 
-    supports_packed_decode: bool = not is_cpu() and not is_npu()
+    # XPU has no tvm_ffi CUDA JIT kernel for KDA packed decode; route XPU to the
+    # non-packed Triton decode() path (fused_sigmoid_gating_delta_rule_update),
+    # the same fallback CPU/NPU use. Batched decode is handled via query_start_loc.
+    supports_packed_decode: bool = not is_cpu() and not is_npu() and not is_xpu()
 
     def packed_decode(
         self,

python/sglang/srt/layers/moe/topk.py

@@ -1741,7 +1741,8 @@ def biased_grouped_topk_gpu(
                 topk_indices,
                 gating_output,
                 renormalize,
-                correction_bias,
+                # The XPU topk_sigmoid AOT kernel requires an fp32 correction bias.
+                correction_bias.to(torch.float32),
                 scale,
             )
 

python/sglang/srt/models/kimi_linear.py

@@ -49,7 +49,7 @@ from sglang.srt.models.deepseek_v2 import DeepseekV2AttentionMLA as KimiMLAAtten
 from sglang.srt.models.llama import LlamaMLP as KimiMLP
 from sglang.srt.models.transformers import maybe_prefix
 from sglang.srt.runtime_context import get_parallel, get_stream
-from sglang.srt.utils import make_layers
+from sglang.srt.utils import is_xpu, make_layers
 from sglang.srt.utils.common import BumpAllocator, add_prefix, set_weight_attrs
 
 
@@ -550,7 +550,7 @@ class KimiLinearModel(nn.Module):
         else:
             self.embed_tokens = PPMissingLayer()
 
-        self.alt_stream = get_stream("alt")
+        self.alt_stream = None if is_xpu() else get_stream("alt")
 
         self.layers, self.start_layer, self.end_layer = make_layers(
             config.num_hidden_layers,